[blove]

Closes #548

Test-only. No product code changes; no changeset needed.

Mechanism (one per assertion, all the same class)

Every eviction-* file settled a window slide with act(() => setTimeout(20)). setRows lands synchronously in the layout effect, but the DOM rows come from the renderer-dom row layout controller, which draws a replacement across scheduler hops (MessageChannel macrotasks, row-layout-controller.ts fallbackScheduler). Under CPU starvation those hops outlast a 20ms timer, and the assertion reads a DOM that is one window behind the props.

Captured under load (run 4-1, load 241–270):

eviction-population-change.test.tsx:180  expected <div aria-rowindex="13" … data-pretable-row-id="row-1"> to be null

row-1 is still rendered from the previous window, but already announced at the NEW window's aria-rowindex (10 + 1 + 2) — the surface has published the new offset while the controller has not yet published the new rows. This is the "one render where a pager swap has moved the window but setRows has not yet settled" skew documented at pretable-surface.tsx (selectionWindow comment), caught mid-flight.

test	assertion	received	verdict
population-change › paints nothing rather than rows the reader never selected	:180 row-1 null	row-1 still rendered (aria-rowindex=13)	one-shot on controller settle
population-change › still paints the selection back…	:229 paint report	rows row-10..row-19 (previous window), nothing selected	one-shot
gate-blip › keeps selection + cursor across an estimated total	:179 after	[] — DOM still at rows 10..19	one-shot
gate-blip › LOCAL MODE	:216 cursor []	["row-8","name"] — row-8 not yet removed from DOM	one-shot
selection-surface › survives an incremental slide…	:173 row-1 null	row-1 still rendered (aria-rowindex=8)	one-shot

No real race. Traced the eviction → selection → paint flow to rule out a product bug:

• create-grid-ui-core.ts observeRowModelRevision: a controller observation whose revision no longer matches the model's current snapshot is skipped (capturedRevision !== revision), so a lagging intermediate revision is never reconciled against a newer window (the "chimera" the getWindowing comment warns about cannot happen through this path).
• The transient render that draws new rows under the pre-reconcile selection fails closed: indexedRangeContainsCell → rangeDatasetSpan → spanReadableInWindow refuses a span whose datasetTotal differs from the window's, so a prepended population paints nothing even for one frame.
• Every captured failure had zero rows painted selected that were not selected.

Deterministic reproduction without load

Temporarily replacing the controller's scheduler hop with setTimeout(task, 30) (not committed): 5 of 11 eviction tests failed, including all three of the issue's cluster members, with identical DOM-lag shapes. With the poll in place, 11/11 under the same injected delay.

Fix

Poll for what the controller drew (waitFor, 15s, per-test 20s timeouts like the sibling indexed-rendering file):

• settledRows(container, ids) — exact row ids in DOM order. Once they match, the same commit has already run observeRowModelRevision, so the paint read next is the settled one.
• gate-blip settledWindow(container, ids, spacerRows) — ids plus the leading spacer the plan placed the first row under (rows[].top is global). A shut gate draws 0 spacer rows, an open one windowStart — this is the only observable difference between the two gates, and it pins that a shut gate really draws no spacer.
• gate-blip's "gate restored, same window" step gets no wait, because it never reached the engine: identical rows are not an effective setRows (create-local-row-model.ts bumps the revision only when drafted.effective) and the controller deliberately does not replan on a resultMeta-only change (pretable-surface.tsx, windowGap comment). The old 20ms sleep there was waiting for nothing; the test's real reconciles are the blip revision and the return slide, both polled.

No claim was weakened: the negative "paints nothing" assertion is read only after the positive ids poll settles, so it cannot pass on a stale DOM.

Mutation results (poll in place, each reverted)

mutation	result
A: spanReadableInWindow fails open on datasetTotal	paints nothing rather than rows the reader never selected RED — new-1..new-4, row-0 painted selected
B: observeRowModelRevision reads a shut gate as local mode (windowed: selectionWindow !== null)	both shut-gate tests RED (after: []), CONTROL + LOCAL MODE green
C: spanReadableInWindow always false	survives an incremental slide… then extends RED (extended span undefined), plus two other span readers

Loaded-run tallies (full @pretable/react suite, 3 concurrent + 10 yes spinners, 10-core Mac)

• Before: 12 runs, 1-min load 138–270. 2 failed, both eviction-population-change (run 2-2 still paints the selection back…, run 4-1 paints nothing…). 1778 tests/run.
• After: 12 runs, 1-min load 195–227, same method. 0 failed — 1778/1778 in every run.

Also under load, no file other than eviction-* failed in either series.

Residual (not changed here)

rejected-rows-data-honesty.test.tsx:597 uses the same act(setTimeout(20)) settle. It did not fail in the issue's runs or in these 24, so it is left alone; it is the same class if it ever does.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
pretable	Ignored		Sep 3, 2026 3:15pm UTC

[github-actions]

Vercel preview ready

Preview: https://pretable-7b00l7r9f-cacheplane.vercel.app
Commit: 444fe4834d4f42d6fb6604920ac848a4c3d0f1c3

_{Updated automatically by the deploy-preview job.}

[blove]

Review fixes — 444fe483

1. gate-blip restore step — comment corrected; the processing-blip restore is now awaited (settledWindow(ids(10), 10): rows 10..19 under the reopened gate's 10-row spacer) via a restoreSpacerRows: number | null param, so the observed-N+2 reconcile is exercised every run. Total blip and CONTROL pass null — verified nothing reaches the controller there (identical rows → drafted.effective false → no revision; no replan on resultMeta-only change).
2. rejected-rows-data-honesty:597 converted, per leg by observable: keep-window → none (said at the site, unawaited); control → leading spacer collapses to 0 rows (settledWindow(SECOND, 0)); refused → no DOM observable — a rejected write retains the coherent window so the spacer stays at 10 rows (the sibling test pins 102..104) — so it polls telemetry selectedRowId to the control leg's settled verdict (pinned null), keeping the differential answer-agnostic. Also checked controller.setViewport on queryChanged: it republishes a stale-snapshot plan only when viewport values change, and they don't here, so the spacer collapse cannot precede the revision landing. Mutation: forcing the reconcile to windowed: true → control leg red at toEqual([]), full selection received.
3. Helpers hoisted to packages/react/src/__tests__/window-settle.ts (flat-module convention of the directory, like rejected-write-harness.ts); all four files import it.

Verification: 4 files normally → 21/21; lint, typecheck, prettier clean. Loaded pass (3 concurrent full suites + 10 spinners, 1-min load 176): 3/3 × 1778 passed.

Spacer-staleness behavior untouched, per instruction.